Doctoral Dissertation ANN-based Short-term Load Forecasting for Rolling horizon Operation of Microgrids Department of Electrical Engineering Graduate School, Chonnam National University NGO MINH DUC February 2019 CONTENTS Pages CONTENTS………………………………………………………………………….i LIST OF FIGURES……………………………………………………………….iv LIST OF TABLES…………………………………………………………………. BACKGROUND AND LITERATURE REVIEW……………………….1 Introduction to microgrid system………………………………………………4 2.1 Each part of microgrid energy management system……………………………5 2.2 Application of microgrid in building sector…………………………………….2 Load forecasting method overview.1 Long-term load forecasting…………………………………………………….2 Medium-term load forecasting………………………………………………….3 Short-term load forecasting…………………………………………………….3 Literature review of short-term load forecasting system………………….2 Machine learning based approaches…………………………………………. DATA PREPARERATION AND ANALYSIS……………………………….2 Characteristic of load profile………………………………………………….3 Correlation between outdoor factors……………………………………………29 3. DATA FEATURES AND GENERATION…………………………………….3 Previous load features……………………………………………………………39 4.
DESIGN AND IMPLEMENTATION STRATEGY OF LOAD FORECASTING SYSTEM FOR MICROGRID ESS SCHEDULING.2 Sparse clustering algorithm…………………………………………………….4 Design and implementation strategy of load forecasting system for microgrid ESS scheduling scheme……………………………………………………………. EXPERIMENTAL RESULTS AND DISCUSSION………………………….3 Comparison with other prediction methods…………………………………….4 Impact of temperature……………………………………………………………63 6.5 Impact of binary features……………………………………………………….6 The impact of different training methods……………………………………….7 Evaluate performance of implementation strategies for microgrid ESS scheduling………………………………………………….8 Analysis of forecasting error for rolling horizon operation approach………….9 Length evaluation of training data………………………………………………71 6.85 iii LIST OF FIGURES Figure Content Page Fig. Microgrid platform 4 Fig. Illustration of typical microgrid energy management system platform………………………………………………………… 5 Fig.
Data preparation process………………………………………. Illustration of impute missing data based on different impute approach………………………………………………………… 19 Fig. Identification of outliner position based on Gradient…………. Flowchart of proposed preprocessing data 22 Fig.
Proposed method for removing outliner and missing data……. Daily energy consumption variation for two years……………. Distribution of energy consumption for each month…………… 26 Fig. Distribution of energy use for day of the week………………… 27 Fig.
Characteristic of weekly load pattern…………………………. Distribution of hourly energy consumption……………………. Daily average temperature curve………………………………. Correlation between outdoor temperature and energy consumption for a university campus building………………… 30 Fig.
Scatter plot between outdoor temperature and energy consumption at each hour for a university campus building…… 32 Fig. Daily average humidity curve…………………………………. Correlation of humidity with energy consumption……………. Daily average wind speed curve……………………………….
Correlation of wind speed with energy consumption…………. Illustration of clustering technique……………………………. Inputs-outputs forecasting model……………………………… 48 Fig. The proposed load forecasting scheme….
One week load pattern with different method…………………. Correlation between real and predicted value…………………. One day load pattern with different methods…………………. ANN- Kmeans with different training methods……………….
ANN- Kmeans with different training methods………………. MAPE distribution vs methods for building 6…………………. MAPE distribution vs methods for 4 buildings …………………. Overestimate MAPE distribution of Building 6………………… 69 Fig.
Overestimate MAPE distribution of StudenHall………………. Overestimate MAPE distribution of HighSchool………………. Overestimate MAPE distribution of Building3_5……………… 71 v LIST OF TABLES Table Content Page Table 1. Replacing missing data and outliers………………………….
Statistics of weather variables dataset and relation with energy consumption………………………………………………… 36 Table 3. Summary of selected features in 24 hours…………………. Basic K-means algorithm…………………………………… 44 Table 5. The algorithm of Sparsified K-means……………………….
The performance one week of the proposed approach and other methods………………………………………………. The performance one day of the proposed approach and other method………………………………………………………. Impact of temperature to forecasting accuracy……………… 63 Table 9. Impact of binary features to forecasting accuracy…………… 64 Table 10.
MAPE of different training methods………………………… 65 Table 11. Comparison between different implementation strategies in Building 6…………………………………………………… 66 Table 12. Comparison between different implementation strategies in StudentHall…………………………………………………… 67 Table 13. Comparison between different implementation strategies in HighSchool…………………………………………………… 67 Table 14.
Comparison between different implementation strategies in Building3_5…………………………………………………… 68 vi Table 15. Length evaluation of training data……………………………. The computation cost…………………………………………. 73 vii NOMENCLARTURE AI Artificial Intelligence ANN Artificial neural network AR Auto regressive model ARMA Auto regressive moving average model ARIMA Auto regressive integrated moving average model ARIMAX Autoregressive integrated moving average with exogeneous model ADN Active Distribution Network DG Distributed generation DBSCAN Density-based clustering algorithm DNN Deep neural networks DERs Distributed energy resources EMS Energy management system ESS Energy storage system FS Feature selection LR Linear regression LMBP Levenberg-Marquardt back propagation LTLF Long-term load forecasting MA Moving average model MLR Multiple linear regression MG MicroGrid viii MAE Mean absolute error MAPE Mean absolute percentage error ML Machine learning MSE Mean square error MTLF Medium-term load forecasting PEIs Power electronic interfaces PCA Principal component analysis RT Regression tree RBP Resilient back propagation RMSE Root Mean Square Error SVR Support Vector Regression SVM Support Vector Machine STLF Short-term load forecasting ix ANN-based Short-term Load Forecasting for Rolling horizon Operation of Microgrids NGO MINH DUC Department of Electrical Engineering Graduate School, Chonnam National University (Supervised by Professor AHN, Seon-Ju) (Abstract) Distributed generation has rapidly grown in recent years as an alternative to traditional centralized electricity generation.
This is mainly driven by the potential advantages of distributed energy resources, such as lower emission, higher conversion efficiency, and the potential to lower the costs and losses in the transmission and distribution system. To realize the advantages of distributed generation, a micorgrid concept is proposed, where generation and associated loads are considered as a subsystem. Short-term load forecast (STLF) is a key input to microgrid energy management system (EMS) for optimal operation of a microgrid. Load forecasting in a short-term is very challenging and complex since the power demand depends on several factors, such as the day of the week, month of the year, temperature, etc.
These difficulties are more complicated when the microgrid architecture changes from large-scale to small-scale, especially at the building level. At this level, forecasting the energy consumption is more challenging due to some problems. Building load profile is small and strongly nonlinear. This makes difficult in getting 1 high accuracy of forecasting.
The missing data and outliers is also raising as an important problem and needed to solve. This thesis studies the STLF for a microgrid using the artificial neural network (ANN) combined with the clustering technique. Firstly, the input features are selected based on the data analysis. Then the model structure and the number of neurons in a hidden-layer is determined.
The performance of the proposed ANN based approach and the conventional methods are compared. To implement the ANN based STLF approach, several methods are proposed for the data preprocessing. First, an algorithm to treat missing and outlier data is proposed. In this method, the gradient is used to detect the outlier data and a clustering based approach is used to replace the missing and outlier data with the approximated value.
Second, a method to automatically assign the binary feature values is proposed. This can significantly reduce human intervention and increase accuracy, especially in case of data variation. Energy storage system (ESS) is one of the essential components of microgrids, which plays an important role in energy management and power trading or sharing. To deal with the forecast uncertainty and to improve the operational efficiency, rolling-horizon strategy can be adopted for the ESS scheduling.
In this approach, the ESS scheduled is determined every hour by using the latest system information. For this operation scheme, STLF should also be run every hour. Therefore, this thesis proposes an implementation process of ANN based STLF for the rolling-horizon concept. To do this, the load data of the last hour is updated and the model training process is repeated every hour.
All the above proposed solutions have been evaluated with actual data of four buildings in our university. The simulation test shows that the proposed rolling-horizon approach with periodical model update can improved the forecast accuracy. The average mean absolute percentage error (MAPE) and the standard deviation of the MAPE is significantly reduced for all four buildings. Introduction Recently, the cutting-edge technologies in smart grids and the fast increasing rate at renewable energy resources field.
These factors induced a lot of new perspectives as well as challenges for managing and controlling energy in power system. This becomes more important when the development trend transforms from large scale (national wide) to small scale sectors like building microgrids. According to the many surveys, building sector is responsible for large portion of total energy use (approximately 40%) and still continue growing in future. Effectively managing energy in this sector will significantly contribute in energy saving.
Therefore energy management system at buildings sector has been received much intention from researchers. The challenges in energy management and power trading in building microgrids become more complicated due to some reasons: Firstly, forecasting energy consumption of small buildings, which contain highly fluctuating characteristic than a large nationwide load, is more challenging. Secondly, How to deal with missing and outliers data, which do not appear at some points but also long patters or few weeks. The high accuracy of forecasting is essential in enhancing the efficiency of the microgrid energy management system.
In the literature, considerable research works have been done for this aspect. One approach follows the conventional method and the other direction follows the modern approach. Conventional methods contain a lot of disadvantages since requiring deeply statistical analysis. In this study, we deploy a modern methodology combined with a clustering method to overcome challenges in forecasting at the small level.
Especially, in the system with dominant ESS, we design a method focusing on rolling horizon operation of ESS. For enhancing the ESS scheduling operation based on rolling-horizon, which depends much on the latest information of load. We proposed an implementation strategy with hourly update model and ensure the well running with rolling-horizon scheduling of ESS. Background and literature review 2.1 Introduction to microgrid system Microgrids are small-scale subsystem with combining of DERs, small generators, loads, electrical vehicles, etc.
The purposed design of microgrids is supplying both electricity and heat for local areas [1]. Microgrids can flexibly operate at both connected and islanded mode. Nowadays, the lack of conventional energy resources from nature such as coal, oil and gas has led the world faced with the crisis of energy many environmental problems. The electrical power system field is not out of this context due to the traditional power system centrally generated power as well as rely on much on these natural resources.
Microgrid concept is the key solution for addressing technical, economic, and environmental issues of modern electrical power system. The rapid development of microgrid paradigm is an important factor in transforming passive distribution network into active distribution network (ADN). In an active distribution network mechanism, the energy is dominantly distributed generated from renewable energy resources such as solar energy, wind energy. 4 In this new context, the energy management energy system (EMS) is also facing more difficulties, which come from the uncertainty of renewable energy resources and from deregulated electricity markets.
These difficulties are more complicated when the microgrid architecture changes from large-scale to small-scale, especially at the building level.1 Each part of microgrid energy management system Grid side Microgrid EMS Customer side Utility PV Forecast Optimization interface interface Wind Weather forcast CHP Data analysis HMI ESS 3rd Pro- consumer EV. Illustration of typical microgrid energy management system platform. MicroGrid (MG) technology has been proposed as one of the most critical solutions for energy poverty problem. In MG technology, MicroGrid Energy Management System (MG- EMS) takes a vital role in control, monitoring and optimizing the operation of the system.
There are two general approaches: real-time immediate technique and predictive technique to meet operational objectives such as cost, power sharing, and so on. The later approach for operating of MG-EMS is more popular, therefore STLF is one of the crucial components in EMS platform. It provides input information from both the grid side and customer side for optimization component in order to archive the optimal operation for system. In this point of view, the implementation forecasting process should consistently go along with operation scheduling scheme.2 Application of microgrid in building sector Based on surveys of many researchers, buildings are one of the fastest growing consuming sectors and the demand for energy of buildings will continue to grow.
It is estimated that the amount of energy consumed by buildings in 2015 accounted for nearly 41% of the total energy use in European Union’s and 73% of total electricity generated in the United States [2-3]. In Korea, referring to reference [4] which has been surveyed every three years, the commercial building sector accounted for 30%-40% of total power consumption [5].